Prostate cancer is the most commonly diagnosed cancer in men. About 25% of patients choose radiation therapy as their primary treatment. Modern radiotherapy aims to deliver a precisely targeted dose to the prostate while protecting nearby healthy structures such as the rectum, bladder, and femoral heads.
Computed tomography (CT) is the standard imaging modality for radiation treatment planning because it provides the tissue density information -- measured in Hounsfield units (HU) -- needed to calculate how radiation will travel through and interact with the body. However, CT offers poor contrast between the prostate, rectum, and surrounding muscles, which can lead to overestimating the prostate volume by up to 30%.
Magnetic resonance imaging (MRI) offers far superior soft tissue contrast and is the preferred modality for accurately delineating the prostate and nearby structures. However, MRI images cannot directly provide the density values needed for dose calculation. As a result, most centers require both a CT and an MRI, which must be carefully aligned with each other before treatment planning can proceed.
Aligning CT and MRI images introduces registration uncertainty -- typically 2 to 5 mm -- that is a systematic source of error persisting throughout the entire treatment course. An MRI-only workflow that eliminates the CT entirely would remove this error source, reduce the patient's radiation exposure from the CT scan, and simplify the planning process.
Proton beam therapy, specifically pencil beam scanning (PBS), delivers protons spot-by-spot with precise control over beam intensity, position, and depth. A proton beam deposits most of its energy at a specific depth -- the Bragg peak -- and delivers very little dose beyond that point. This spares tissues on the far side of the tumor more effectively than X-ray radiation.
However, because proton dose depends on the stopping power of tissues along the beam path, it is even more sensitive to tissue density errors than conventional radiation. Protons must traverse large pelvic bony structures like the femoral heads and pubic symphysis before reaching the prostate, making accurate bone HU prediction critically important.
To enable an MRI-only workflow, a synthetic CT (sCT) image must be generated from the MRI. This sCT serves as a CT substitute, providing the HU values needed for dose calculation and for generating reference images used in daily patient setup. The challenge is that bones -- which have a distinctive bright appearance on CT -- appear ambiguous on MRI, making their HU values difficult to predict accurately.
Existing sCT approaches fall into three categories: segmentation-based methods (which assign bulk densities to tissue types), atlas-based methods (which deform a template CT to match the patient's MRI), and machine learning-based methods (which learn the MRI-to-CT mapping directly from patient data pairs). Machine learning offers the most flexible and patient-specific predictions but requires careful validation.
The study tested a deep learning approach based on a 3D cycle-consistent generative adversarial network (cycleGAN) enhanced with dense blocks. Standard GAN-based methods struggle when MRI and CT images are not perfectly aligned, producing blurred predictions at tissue boundaries. CycleGAN addresses this by learning a two-way translation: MRI to CT and CT back to MRI, enforcing internal consistency even with imperfectly matched image pairs.
Dense blocks connect each layer of the network to all subsequent layers, allowing the network to carry forward both fine detail (high-frequency information) and broad context (low-frequency information) throughout the architecture. This enables more accurate reconstruction of both soft tissue and bone boundaries, which is critical for proton dose calculation.
The network was trained on 17 prostate cancer patients using leave-one-out cross-validation -- each patient's sCT was generated using a model trained on all the other patients. The MRI and CT images were co-registered and matched in resolution before training. Training required about 15 hours on a high-end GPU, but generating the sCT for a new patient takes only about 2 minutes.
The loss function guiding training combined an adversarial loss (encouraging realistic-looking CT images) with a distance loss that penalizes both intensity errors and boundary blurring. The distance used a fractional norm (lp with p=1.5) that avoids the blurring caused by mean squared error while also avoiding the edge artifacts of mean absolute error.
Image quality was assessed using three standard metrics: mean absolute error (MAE) comparing HU values between the sCT and the reference CT, peak signal-to-noise ratio (PSNR), and normalized cross-correlation (NCC). Bone region quality was separately evaluated because accurate bone density prediction is especially critical for proton planning.
Dosimetric accuracy was assessed by designing actual proton treatment plans on both the original CT and the sCT using the same beam arrangements, then comparing the resulting dose distributions. Metrics included dose-volume histogram (DVH) comparisons for the tumor target and all nearby organs at risk -- the rectum, bladder, femoral heads, penile bulb, and seminal vesicles.
Gamma analysis was used to quantify agreement between dose distributions in three anatomical planes. This standard clinical quality assurance test measures how many points in the dose map pass both a dose tolerance (1%, 2%, or 3%) and a spatial tolerance (1 mm, 2 mm, or 3 mm) simultaneously. Higher pass rates indicate better agreement.
Two proton-specific tests were also applied. The distal range analysis compared the depth at which each proton beam deposited 80% of its peak dose, checking that sCT-based plans fell within the established clinical tolerance criterion. An innovative individual pencil beam Bragg peak shift analysis tracked the spatial displacement of each individual beam's energy deposition point between CT-based and sCT-based plans -- a more sensitive and comprehensive measure than the range analysis alone.
The proposed method achieved an average mean absolute error (MAE) of 51.32 HU across the whole image, comparable to state-of-the-art methods. The bone region, which is most critical for proton planning, had a bone MAE of about 105 HU -- substantially better than the competing deep CNN method (141 HU) and GAN method (130 HU) tested on the same patients.
Dosimetric agreement between CT-based and sCT-based plans was strong. The mean dose difference was just -0.07%, and the mean absolute dose difference was 0.23% of the prescribed dose. Gamma analysis pass rates were 92.4% at the strictest 1 mm/1% criterion and exceeded 98% at the 3 mm/3% criterion -- all within clinically acceptable ranges.
For the tumor target (PTV), DVH metric differences were generally under 1%. Organs at risk showed mostly small differences, though the rectum and bladder showed larger variability in a few patients. The main reasons were differences in organ filling between the separately acquired MRI and CT scans, and the challenge of accurately predicting HU values in bone regions near those organs.
All 17 patients had proton beam range displacements within the established MGH clinical tolerance criterion, confirming that the sCT-based plans would be safe to use. The median Bragg peak shift across all individual beam spots was just 0.09 cm -- a clinically negligible displacement for most spots, with only isolated large outliers at tissue-air interfaces.
The largest source of dose disagreement was at the distal edge of proton beams, particularly where beams crossed bony structures. Bones appear ambiguous on MRI -- they can look dark or have variable signal depending on the scan sequence -- which makes their HU values hard to predict accurately. Even small bone density errors translate into range shifts that are noticeable in proton dose distributions.
One technical limitation is that deep learning networks typically apply intensity normalization that slightly clips the range of predicted HU values. This causes a subtle underestimation of bone HU, which leads to the proton beam traveling slightly deeper than planned. Addressing this is important for future clinical implementation.
A practical constraint of MRI scanners is their limited field of view (FOV), which often does not extend to cover the full pelvis laterally. In this study, tissue outside the MRI FOV was supplemented from the original CT, meaning the evaluation was partial. Future use of large-bore MRI scanners, which are already available in some centers, would resolve this limitation.
An important finding was that standard image quality metrics (MAE, PSNR, NCC) were not well correlated with actual treatment plan accuracy metrics such as gamma pass rate, range shift, and Bragg peak shift. This means a good-looking sCT image does not automatically guarantee a clinically accurate treatment plan, and proton-specific dosimetric evaluation should always accompany sCT validation studies.
This study demonstrates the feasibility of MRI-only proton treatment planning for prostate cancer using a 3D dense cycleGAN approach. The dosimetric accuracy of sCT-based plans was comparable to CT-based plans across all evaluated metrics, supporting progression to further clinical development and prospective evaluation.
The study also introduced and validated the individual pencil beam Bragg peak shift analysis as a comprehensive new evaluation tool for sCT quality in proton planning. Unlike global range measures or dose-plane gamma analysis, this metric directly tracks each spot's energy deposition point, providing a more sensitive and spatially resolved picture of sCT accuracy.
Future directions for MRI-only proton planning include moving beyond HU-based dose calculation toward direct prediction of the stopping power map from MRI, which is the quantity proton dose calculation actually needs. Generating tissue elemental composition maps for use in Monte Carlo simulations is another promising avenue that would further reduce the systematic uncertainties in proton treatment planning.
The trend toward MRI-guided radiation therapy -- using real-time MRI during treatment delivery -- adds further motivation for this work. If sCT images can be generated quickly and accurately from MRI, the same imaging used to guide daily treatment could also be used for adaptive online replanning, opening the door to more precise and personalized proton therapy.